• 해당 논문은14th ICAEIC-2025 에서 발표되었습니다.
Optimized Long Short-Term Memory Deep Learning Approach for Precise Software Reliability Fault Estimation
Juwon Jung1, and Chaebong Sohn2
모아소프트㈜
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Key Words : LSTM (Long Short-Term Memory), Prediction Techniques, Software Reliability Growth Models
(SRGM), Reliability Assessment, Time Series Analysis
서론
This study proposes an LSTM (Long Short-Term Memory)-based deep learning model to overcome the
concerns of existing Software Reliability Growth Models (SRGM), which suffer from reduced prediction
accuracy due to limited fault data in the early software testing environment. The study involves generating
synthetic fault data for the software testing phase, optimizing model performance through hyperparameter
tuning with Bayesian Optimization, and applying GPU memory optimization and a sliding window technique.
The results demonstrate that the proposed LSTM model outperforms traditional SRGM in prediction accuracy.
In the 1–60 time interval, the RMSE of the LSTM model is approximately 37, which is 214 lower than the
RMSE of 251 observed for SRGM. Notably, the LSTM model exhibited stable performance in both the early
stages with insufficient data and in long-term predictions beyond 61 time intervals. The LSTM-based deep
learning model presented in this study offers a novel alternative for software reliability prediction by deeply
learning complex long-term patterns in time-series data and combining this capability with optimization
methodologies that explore global minimum. This approach enhances the model’s generalization performance
and ensures consistent results under various initial conditions. Such a technical approach is expected to be
highly applicable in advanced industries that demand high-reliability systems, such as defense, aerospace,
and healthcare.
…중략
결론
This study proposes an LSTM-based deep learning model for fault prediction in the software testing phase and compares and analyzes its performance with the existing SRGM model. The research results confirmed
that the LSTM-based model demonstrated superior prediction accuracy compared to traditional SRGM, even
in situations with limited fault data. Key findings of this study are as follows.
-Prediction Accuracy: The LSTM model achieved higher prediction accuracy, demonstrating an
RMSE 214 lower than the SRGM model in the 1-60 time interval.
– Response to Data Scarcity: Even in environments with limited data, the LSTM model exhibited more
stable and accurate predictive performance compared to the SRGM model.
– Long-term Prediction Capability: The LSTM model effectively predicted trends reflecting existing
fault data, even in intervals beyond the 61st time point where actual data was unavailable.
-Model Optimization: Model performance was maximized through hyperparameter tuning using
Bayesian optimization techniques.
-Data Utilization Efficiency: The application of the sliding window technique and GPU memory
increase enabled effective learning even with limited datasets.
These results suggest that the LSTM-based deep learning model has significant advantages in learning long
term and complicated patterns of time series data and predicting future fault occurrences in software reliability analysis. The model effectively complements the concerns of SRGM, particularly in situations where data is scarce or incomplete.
These results suggest that LSTM-based deep learning models can overcome the concerns of existing SRGM models in the field of software fault prediction, providing more accurate and stable predictions. The LSTM model is particularly effective in situations with limited data or when long-term predictions are required. Future research could further generalize these findings by validating the results using data from various real software projects and conducting comparative analyses with other deep learning models. Additionally, research into methods for improving model interpretability and providing reliability measures for prediction results is necessary.
In conclusion, the LSTM-based deep learning model proposed in this study presents new possibilities in the
field of software reliability prediction. It is expected to be particularly useful in areas requiring high reliability,
such as the defense industry.
- 국방무기체계 SW ALM 도구 관련 문의 : 02-6945-2156 / Contact us – (주)모아소프트 (moasoftware.co.kr)
